System and method for estimating perturbation norm for the spectrum of robustness

Inventors

RICE, Leslie • KOLTER, Jeremy • Lin, Wan-Yi

Abstract

A computer-program product storing instructions which, when executed by a computer, cause the computer to, for one or more iterations, update parameters associated with a machine-learning network utilizing perturbations for input data, wherein the perturbations are sampled utilizing Markov chain Monte Carlo, identify a loss value associated with each perturbation in each iteration, and evaluate the machine learning network by identifying an average loss value across each iteration and outputting the average loss value.

Core Innovation

The invention provides a computer-implemented approach for training a neural network using input data from one or more sensors, including time-series data, image data, video data, or sound data. During training, random perturbation samples are associated with the input data and are used to compute loss values for determining gradients of the loss with respect to neural network parameters. The parameter updates are performed using one or more machine learning optimizers in response to an intermediate-p robustness objective.

The intermediate-p robustness is defined as an expectation of a p-norm of a loss, with 1<p<∞. The training framework increases a level of perturbation stress across successive iterations, applying the increased perturbation stress to new perturbations for one or more successive iterations. A first threshold associated with convergence of the neural network is used to determine when to output a trained neural network.

To support robustness evaluation and optimization without a tractable robustness integral, the invention employs path sampling and Markov chain Monte Carlo to sample perturbations from a perturbation density. The intermediate-p robustness is approximated using a geometric-mean path estimator based on sampled losses. The framework links this intermediate-p robustness evaluation to training by computing loss values per perturbation, deriving gradients of the intermediate-p robustness with respect to neural network parameters, and updating the parameters using gradient-based optimization.

Claims Coverage

The partial claims include three independent claims: a computer-implemented method, a system, and a computer-program product. Across the independent claims, the coverage centers on training a neural network using sensor input data with random perturbations sampled via Markov chain Monte Carlo or derived from path sampling, updating parameters using gradients tied to an intermediate-p robustness objective defined as an expectation of a p-norm of a loss, while increasing perturbation stress and outputting a trained network upon convergence to a first threshold.

Across the independent claims, the coverage is directed to training and evaluation of a neural network using sensor-derived input data with random perturbations, where the optimization target is intermediate-p robustness expressed as an expectation of a p-norm of a loss (1<p<∞). The claims further require increasing perturbation stress over successive iterations, sampling perturbations using Markov chain Monte Carlo and/or deriving perturbation samples from path sampling, computing loss-related gradients tied to the intermediate-p robustness, and outputting a trained model upon convergence to a first threshold.

Stated Advantages

Documented Applications